The rules will arrive finished

On Tuesday 15 September 2026 Ecuador's National Assembly shelved, with 88 votes, the Organic Law for the Regulation and Promotion of Artificial Intelligence, after two years in committee ^[1]. The committee report that recommended it, approved on 31 August, gives the reason in a sentence worth reading whole: a rigid regulation «could cause regulatory obsolescence, barriers to innovation, disincentives to technological investment and heavier burdens on start-ups and small firms» ^[2]. The same report proposes what to do instead of legislating: «consider international benchmarks for the responsible management of artificial intelligence, among them the ISO/IEC 42001:2023 standard» ^[2]. One legislator put it with less diplomacy on the floor, according to the session's chronicle: the right course is «to remain watchful of what the great powers do» ^[1]. The committee's president filed a replacement bill on 1 September, and the country was left, for now, without a law ^[2].

Ecuador is today the best example of a regional movement, and it is worth naming before its protagonists. Latin America did not decide against regulating artificial intelligence: it decided to import the rulebook, and the rulebook is being written, this very month, by those who will be regulated. Chile, Brazil and Ecuador reached the same conclusion by different roads, and only Ecuador names whom it is delegating the decision to.

Three capitals, one sentence

Chile had passed in its Chamber, in October 2025, a framework AI law modelled on Europe's, with four risk categories, prohibitions and sanctions ^[3]. Kast's government wants to replace it entirely. The science minister, Ximena Lincolao, called it «incongruent» before the Senate and said that Europe «has given us lessons in what not to do» ^[4]. Her proposal, presented to senators in May, has as its first pillar, in her own explanation, an enabling regulatory framework «with the possibility of automatic homologation of international certifications», followed by sandboxes, voluntary audits and sanctions «proportional to a concrete harm and not to a theoretical risk» ^{[4] [5]}. The Chilean press described it as «updatable technical standards, certifications and voluntary audits» ^[3]. None of those presentations names which certifications will be homologated or who writes them, and the academics La Tercera consulted warned of the risk of «regulatory capture if the large companies end up shaping a framework designed to fit them» ^[3]. Lincolao promised the text for early September. The promise still has no text: as of 22 September the Chamber's public record shows only that the government renewed the bill's simple urgency on the 1st, and no amendment filed ^[6].

Brazil passed its law in the Senate in December 2024, and the Chamber has held it since March 2025 «awaiting the rapporteur's opinion», with five voting dates missed ^{[7] [8]}. The wait now has a date. The rapporteur, Aguinaldo Ribeiro, said in August that the vote will come after October's elections ^[8]. The text that waits is less autonomous than it looks. Its competent authority must «encourage the adoption of internationally recognised standards, best practices and certifications», its sectoral authorities must promote «harmonisation with national legislation and with international norms», and the bodies that will certify high-risk systems are accredited «in conformity with applicable national and international technical standards» ^[9]. The Brazilian law names neither ISO, nor the OECD, nor NIST: it delegates to «international norms» without saying which, that is, to whoever writes them.

Mexico has no general law and first needs to amend two articles of its Constitution before Congress can pass one ^[10]. President Sheinbaum said so on 14 September at her morning press conference: regulating artificial intelligence and social media «is a pending discussion», «it has its complications, how far, how», and «the electoral period has already begun, I don't know if it can still be done at this moment» ^[11]. Colombia is on its sixth attempt. It shelved the fifth bill in June, at the end of the legislative term and without a first debate, and the sixth, filed in July, states in its explanatory memorandum that it draws on the OECD, UNESCO and the European law ^[12]. Argentina has had a committee opinion since November 2025 and no floor vote, with the governing bloc warning against «technological populism through premature legislation that limits the advance of innovation» ^[13]. The UN's advisory body on artificial intelligence counted in 2024 that 118 countries, «primarily in the global South», belong to none of the seven major governance initiatives; seven countries belong to all of them, and none is Latin American ^[14].

The rulebook they will homologate

While the legislatures wait, the rulebook is being written, and the labs are writing it. On 12 September Dario Amodei proposed that every frontier lab give «employee-like access» to embedded outside evaluators, and asked for a narrow antitrust waiver so as to coordinate with its rivals ^[15]. Six days later Anthropic named its first evaluator, Accenture's Faculty unit, and acknowledged in the announcement that «there are, as yet, no standards for what information embedded evaluators should have access to, or how they should report what they find», nor a system to fund them, so Anthropic pays its own directly ^[16]. The first to arrive is paid by the evaluated. Demis Hassabis had proposed in July an industry-funded body to which models would be submitted up to thirty days before release ^[17]. Elon Musk proposed on 14 September that rivals review one another, «instead of grading your own homework», and admitted that none has agreed ^[18]. The three proposals differ on who judges and agree on who decides the design: in none does a government other than that of the United States take part, and in none does a country that only adopts appear.

The reference measurement is also home-made. On 17 September Anthropic published that Claude now «leads» 26 % of its own research, with a caveat printed underneath: two human raters looking at the same evidence agreed on the exact level 35 % of the time, and «each developer will be tempted to draw the line generously» ^[19]. The coordination already has a docket. Four subscribers sued Anthropic, OpenAI, SpaceXAI and Google on 18 September for agreeing on the pace among themselves, and Cohere's Aidan Gomez said that the rules for this technology «cannot be written by a small group of commercially aligned companies behind an antitrust waiver» ^{[20] [21]}. Carlos Coello, the Mexican who sits on the UN's scientific panel alongside a Chilean and a Brazilian, three of forty members, described it without adjectives: «They are regulating themselves; they evaluate their own risks and that information is confidential, nobody else has access» ^[22]. Those are the facts.

The region already knows what it is to receive a finished rulebook. The Basel Committee was founded by the central-bank governors of the Group of Ten at the end of 1974, published Basel I in 1988 and Basel II in 2004, and Latin America «has generally followed» its recommendations since the first accord ^{[23] [24]}. Brazil and Mexico got seats on the committee in March 2009, and Argentina in June, twenty-one years after the first standard they were applying ^[23]. The three rating agencies the SEC recognised in 1975 rate twenty-six countries of the region today, and a sovereign is investment grade when two of the three say so ^{[25] [26]}. That arrangement had costs, and the region accepted it because it paid: a bank in Lima or Bogotá is supervised under rules that nobody in Lima or Bogotá drafted, and its depositors sleep better for it. Basel, however, was written by supervisors to supervise banks, and the rating agencies are paid by the issuer but are not the issuer. The artificial-intelligence rulebook is written by the labs to evaluate the labs, with evaluators they pay, with access standards they themselves say do not exist, and with an antitrust suit open for the mere fact of having agreed among themselves.

Laura Schertel Mendes chairs the digital-law commission of Brazil's bar association and wrote in May that «whoever defines technical standards defines markets» ^[27]. The question, in her words, is not whether Brazil should regulate but «whether it prefers to absorb, at a later moment, the rules written by other countries», because «not regulating is outsourcing the decision» ^[27]. Loreto Bravo, the Chilean on the UN panel, said it from the other side: «we are going to become users, like a country that adopts something developed by global actors» ^[28].

The counterpoint has to be conceded whole. Ecuador's committee is right that a rigid law on a technology that changes every month ages before it is implemented, and Lincolao is right that Europe regulated first and is now revising its own law. A country that trains no frontier model gains nothing by writing rules for training them, and an international certification may be better than any article a regional congress could draft. Importing is, for an adopter, the rational decision. The problem is not importing but importing without conditions, and that is what the three texts do: none says which certification it accepts, who has to issue it, or what happens if the certifier is paid by the certified.

The conditions exist and are already written, though elsewhere. The bank-outsourcing rules of Brazil, Mexico and Colombia oblige banks to demand independent audit reports from any foreign technology vendor, and Mexico's makes the vendor accept the supervisor's visits unconditionally ^[29]. A banking supervisor who treats a model's API as a relevant outsourced service regulates adoption without touching development, which is exactly what the region can regulate.

THE PLAY

1. Regulate adoption, not development. Whoever writes the contract with a lab writes the only rule that will exist in the region this year. A banking supervisor, a public-procurement authority or a health ministry can demand today the outside evaluator's report, incident notification and the independence of whoever evaluates, with the rules they already have ^[29].
2. Homologate by name. If Chile is going to homologate «international certifications» automatically, let the law say which ones, who issues them, and that the certifier is not paid by the certified. One line in the substitute bill is worth more than four risk categories, and the same line fits in article 46 of PL 2338 before it is voted.
3. Sit where it is written, not where it is signed. The three Latin American seats on the UN panel and the committees where the standards Ecuador just cited are drafted cost less than a law and arrive before it. A region that is going to import the rulebook has to be, at the very least, in the room where it is corrected.